How do the Realtime API and Kawl Now compare?
The Realtime API is a building block, and Kawl Now is a finished product that uses OpenAI underneath. The table shows what each gives you, from OpenAI’s documentation and pricing as checked on 8 October 2026.
| Consideration | OpenAI Realtime API | Kawl Now |
|---|---|---|
| What it is | An API for building speech-to-speech voice agents. | A ready-made voice agent for SaaS websites. |
| Voice model | gpt-realtime-2.1, its mini version or another Realtime model. | OpenAI’s GPT-Live, chosen by Kawl Now. |
| In the browser | WebRTC, through your server or a short-lived client secret it creates. | One script tag. The embed holds only a public workspace ID. |
| What your team builds | The session server, access checks and limits, call screen, document search and transcript storage. | Approve knowledge, set a greeting and voice, add your sites, paste one script tag and test. |
| Knowledge | Function tools or remote MCP servers that you connect. | Approved FAQs, notes and GitHub docs, searched per agent. |
| Phone calls | Incoming calls through a SIP trunk, such as Twilio’s. | No phone number for customers to call. Visitors talk in the browser. |
| Voices | Ten built-in voices, with marin and cedar recommended, plus custom voices for eligible customers. | Two voices: Marin and Cedar. |
| Longest session | 60 minutes. | 10 minutes. |
| Pricing | Per token: $32 per million audio input tokens and $64 per million audio output tokens on gpt-realtime-2.1. | First 50 conversations free per workspace, a one-time allowance. Paid calls aren’t available yet. |
What is the OpenAI Realtime API?
The Realtime API runs a live session with a speech-to-speech model. OpenAI says the model works directly with audio, keeps the conversation’s state and can call tools, which suits agents that need interruptions, fast first audio and natural turn-taking. Its current model, gpt-realtime-2.1, supports configurable reasoning effort and a 128,000-token context window, and a cheaper mini version is available.
OpenAI now offers a second voice API. Its audio guide recommends starting a new conversational voice app with GPT-Live, a full-duplex model that listens while it speaks and passes reasoning and tool use to a separate backend model. The Realtime API remains the choice when you want speech, reasoning and tools in one session. Kawl Now runs on GPT-Live.
What do you build around it for a website?
OpenAI’s WebRTC guide offers two ways to connect a browser. With the unified interface, the page sends its connection offer to your server, which forwards it to OpenAI with your standard API key. With an ephemeral key, your server requests a short-lived client secret and the browser connects to OpenAI directly. Either way, the standard key stays on your server. The same guide’s GPT-Live example says to protect its session endpoint with authentication, authorisation, request limits and HTTPS before other people can use it, and a Realtime API endpoint needs the same care.
The model handles the conversation. Everything around it on a public website is yours to build and run:
- A session endpoint, plus checks on which sites and visitors may start a call.
- Limits on call length, calls at once and spend.
- The call screen: microphone prompt, captions, mute and an end button.
- Search over your docs, connected as function tools or an MCP server.
- Storage and a review screen for transcripts.
How much does the Realtime API cost?
The Realtime API charges per token, not per minute. On 8 October 2026, OpenAI listed gpt-realtime-2.1 at $32 per million audio input tokens, $0.40 per million cached input tokens and $64 per million audio output tokens, and the mini model at $10 and $20. Caller audio counts as one token per 100 milliseconds and the model’s speech as one per 50 milliseconds, so a minute of the agent speaking is 1,200 output tokens, about $0.077 at the full model’s rate.
Each response sends the whole conversation so far as input, so later turns cost more unless that input is cached. GPT-Live is priced differently, at $0.05 a minute of session time billed per second, with the backend model’s usage on top.
Kawl Now counts conversations instead. Each workspace gets its first 50 free, tests included, and paid calls aren’t available yet, so no payment is collected.
When is building on the Realtime API the better choice?
Build it yourself when voice is part of what you sell, or the agent has to do more than answer questions. The Realtime API fits better if you need:
- A voice assistant inside your own app, with your own interface and behaviour.
- Account data or actions, through function tools that call your systems.
- Phone calls, through a SIP trunk pointed at OpenAI, with transfers by SIP REFER.
- Control of the model, voice, turn detection and reasoning effort for each session.
- Sessions longer than 10 minutes, up to OpenAI’s 60-minute limit.
When does Kawl Now make more sense?
Kawl Now suits a SaaS team that wants voice answers on its website without starting a voice engineering project. It has no phone number for customers to call, no model choice and no custom tools, and it greets callers in English. In return, it gives you:
- Separate support and sales agents, each with its own knowledge, greeting, voice and widget.
- Search limited to sources your team has approved, with GitHub docs reviewed again after every sync.
- Call limits built in: 10 minutes per call, three at once and 12 new calls a minute per workspace.
- Two-sided text transcripts in the dashboard. Kawl Now doesn’t record call audio.
- A REST API and an MCP server, where adding a source never approves it.
Common questions
What is the difference between the OpenAI Realtime API and GPT-Live?
Both are OpenAI voice APIs. The Realtime API runs one speech-to-speech model that hears, reasons, uses tools and speaks in a single session, priced per token. GPT-Live is a full-duplex model that listens while it speaks and hands reasoning and tools to a backend model, priced at $0.05 a minute plus that backend’s usage. OpenAI recommends GPT-Live for new conversational voice apps.
Can you use the OpenAI Realtime API on a website without a server?
Not safely. OpenAI says standard API keys belong on your server, which either forwards the browser’s connection offer or creates a short-lived client secret for it. Kawl Now runs that server for you, so its embed holds only a public workspace ID.
Does Kawl Now use the OpenAI Realtime API?
No. Kawl Now uses OpenAI’s GPT-Live over WebRTC, so OpenAI processes the voice conversation. Kawl Now adds the widget, the session server, search over approved knowledge, call limits and transcript review around it.